Sometimes kernel launch ended with
“unspecified launch failure” error. OS is WinXP x86 SP2.
I have suspictions that it’s GPU hardware-related problem (this GPU, 9400GT had problems in another host/OS too).
But does some way exist to establish if GPU broken or something wrong with OS config or length of kernel itself (for example) ?
An ‘unspecified launch failure’ is generally a segfault (in Unix-speak, I don’t know the equivalent in Windows). Basically something tried writing to a bit of memory it didn’t own. Of course, if you have a flaky card, then I imagine that could cause random trouble too - the only way to test that is to get another card.
THere was a famous ULF thread where people have seen random failures that are un-accouted for. If you are having a pre-release card (mostly univs have access to pre-release cards), then this could be a card problem. Because pre-production cards are pre-production for a reason (borrowing words of tmurray here).
MrAnderson has considerable knowledge on it… He even worked out that compute 1.1 cards or so have this ULF problem… I am not sure… I only remember vaguely him saying so… I dont know if it was ascertained by NV or so…
I would test however if your code really has no bugs before blaiming the “poor” GPU ;)
Some read-after-write or vice-versa hazards or some other strange behaviour are hard to catch.
In parallel environment a program may work correctly million times and fail in this 1000001st case…